PROGNOSTIC INDEX SCORE AND CLINICAL PREDICTION MODEL OF LOCAL REGIONAL RECURRENCE AFTER MASTECTOMY IN BREAST CANCER PATIENTS

Transcription

1 doi: /j.ijrobp Int. J. Radiation Oncology Biol. Phys., Vol. 64, No. 5, pp , 2006 Copyright 2006 Elsevier Inc. Printed in the USA. All rights reserved /06/$ see front matter CLINICAL INVESTIGATION Breast PROGNOSTIC INDEX SCORE AND CLINICAL PREDICTION MODEL OF LOCAL REGIONAL RECURRENCE AFTER MASTECTOMY IN BREAST CANCER PATIENTS SKYE HONGIUN CHENG, M.D.,* # CHENG-FANG HORNG, M.S., JENNIFER L. CLARKE, PH.D., MEI-HUA TSOU, M.D., STELLA Y. TSAI, M.D.,* # CHII-MING CHEN, M.D., JAMES J. JIAN, M.D.,* # MEI-CHIN LIU, M.D., MIKE WEST, PH.D., ANDREW T. HUANG, M.D., ** AND LEONARD R. PROSNITZ, M.D. # Departments of *Radiation Oncology, Research, Pathology, Medical Oncology, and Surgical Oncology, Koo Foundation Sun Yat-Sen Cancer Center, Taipei, Taiwan; Institute of Statistics and Decision Sciences and Departments of # Radiation Oncology, **Medicine, and Biostatistics and Bioinformatics, Duke University, Durham, North Carolina Purpose: To develop clinical prediction models for local regional recurrence (LRR) of breast carcinoma after mastectomy that will be superior to the conventional measures of tumor size and nodal status. Methods and Materials: Clinical information from 1,010 invasive breast cancer patients who had primary modified radical mastectomy formed the database of the training and testing of clinical prognostic and prediction models of LRR. Cox proportional hazards analysis and Bayesian tree analysis were the core methodologies from which these models were built. To generate a prognostic index model, 15 clinical variables were examined for their impact on LRR. Patients were stratified by lymph node involvement (<4 vs. >4) and local regional status (recurrent vs. control) and then, within strata, randomly split into training and test data sets of equal size. To establish prediction tree models, 255 patients were selected by the criteria of having had LRR (53 patients) or no evidence of LRR without postmastectomy radiotherapy (PMRT) (202 patients). Results: With these models, patients can be divided into low-, intermediate-, and high-risk groups on the basis of axillary nodal status, estrogen receptor status, lymphovascular invasion, and age at diagnosis. In the low-risk group, there is no influence of PMRT on either LRR or survival. For intermediate-risk patients, PMRT improves LR control but not metastases-free or overall survival. For the high-risk patients, however, PMRT improves both LR control and metastasis-free and overall survival. Conclusion: The prognostic score and predictive index are useful methods to estimate the risk of LRR in breast cancer patients after mastectomy and for estimating the potential benefits of PMRT. These models provide additional information criteria for selection of patients for PMRT, compared with the traditional selection criteria of nodal status and tumor size Elsevier Inc. Breast cancer, Mastectomy, Radiotherapy, Prediction model, Prognostic score. INTRODUCTION Postmastectomy radiotherapy (PMRT) clearly reduces the frequency of local regional recurrence (LRR) in high-risk breast cancer patients (1). It also seems to impact favorably on survival (2). The delineation of patients at high risk for LRR is controversial, however. Conventionally, the number Note An online CME test for this article can be taken at under Education and Meetings. Reprint requests to: Skye H. Cheng, M.D., Koo Foundation Sun Yat-Sen Cancer Center, Department of Radiation Oncology, No. 125, Lih-Der Road, Pei-Tou District, Taipei, Taiwan. Tel: ( 886) , ext. 310; Fax: ( 886) ; skye@ mail.kfcc.org.tw Supported in part by research funds from Koo Foundation Sun Yat-Sen Cancer Center and in part by a grant from the National Health Research Institutes of Taiwan (Contract Project 1997, No. DD01-86IX-CR601S). Acknowledgments The authors thank the members of the Breast of involved axillary lymph nodes and the size of the primary tumor are considered, and PMRT is generally recommended for those with 4 or more involved axillary lymph nodes and/or those with large primary tumors (T3 or greater) (3 5). However, three randomized trials in which a survival benefit for PMRT was demonstrated included primarily Cancer Team at Koo Foundation Sun Yat-Sen Cancer Center: Drs. Po-Sheng Yang and Ben-Long Yu (Department of Surgery), Drs. H.H. Lin and M.Y. Lee (Department of Pathology), Drs. Kwan- Yee Chan and Christopher K.J. Lin (Department of Radiology), Drs. Tran-Der Tan, Cheng-I Hsieh, and Nei-Min Chu (Department of Medical Oncology), and Dr. Yu-Ling Chung (Department of Radiation Oncology) for patient care; and Yen-Chun Lin, Yueh- Yun Yu, Yi-Wen Chang, and An-Chen Feng in the Clinical Protocol Office for data collection, data entry, data quality control, and outcome analysis. Received April 2, 2005, and in revised form Nov 15, Accepted for publication Nov 15,

2 1402 I. J. Radiation Oncology Biology Physics Volume 64, Number 5, 2006 Table 1. Clinical characteristics of patients in the training and test data sets Table 1. Clinical characteristics of patients in the training and test data sets (Continued) Characteristic Training Test (n 506) (n 504) P value* Characteristic Training Test (n 506) (n 504) P value* Age (y) Menstruation status Premenopausal Postmenopausal Family history No st or 2nd degree Others 3 2 Histology Favorable Infiltrating ductal Other invasive Pathological tumor size (cm) No. of axillary nodes dissected No. of axillary nodes positive Extracapsular extension Negative Positive Unknown 7 5 Estrogen receptor status Negative Lymphovascular invasion Absent Focal Prominent Nuclear grade Surgical margins Negative Close ( 2 mm) 7 14 Positive 1 2 Unknown 5 5 Adjuvant radiotherapy Yes No Adjuvant hormonal therapy Yes No (Continued) Adjuvant chemotherapy Yes No * Chi-square test. Favorable histology: medullary, tubular, or mucinous carcinoma. patients with 1 3 positive lymph nodes (6 8). Thus, a better definition of what constitutes a patient at high risk for LRR and who would be expected to benefit from PMRT would be valuable. Several other prognostic factors, such as estrogen receptor (ER) status, age, lymphovascular invasion (LVI), and extracapsular extension of tumor in axillary lymph nodes have previously been identified as predictive for LRR after mastectomy (9 11). The interaction between these factors, however, is largely unknown. The aim of this study was to develop more sophisticated prediction models for LRR after mastectomy, by using readily available clinical data in a fashion analogous to the development of the International Prognostic Index for patients with non-hodgkin s lymphomas (12). To accomplish this, both traditional Cox proportional hazards models and Bayesian classification trees were used to estimate the probability of LRR after mastectomy for individual breast cancer patients (13 15). METHODS AND MATERIALS Treatment policies Between April 1999 and December 2001, 1,143 patients underwent modified radical mastectomy as initial treatment for newly diagnosed invasive breast cancer at the Koo Foundation, Sun Yat-Sen Cancer Center in Taipei, Taiwan. Adjuvant treatment policies were as follows. Postmastectomy radiotherapy. Before 1997, patients received PMRT if they had involvement of 4 or more axillary lymph nodes or a primary tumor 5 cm in size or a resection margin positive or close. After 1997, patients with 1 3 axillary lymph nodes involved were also candidates for PMRT if combined with another risk factor, specifically a primary tumor 3 cm, ER-negative status, age 40 years, or the presence of LVI. The technique for PMRT included radiation fields specifically directed to the ipsilateral chest wall and internal mammary chain and supraclavicular lymph nodes with CT-based treatment planning. The heart was largely excluded from the radiation fields. The central lung distance of the tangents fields was limited to a maximum of 3 cm. Internal mammary chain nodes were either included in wide tangent fields if the included lung was acceptable or treated with a separate photon/electron field. The full axilla was excluded from the radiation fields. The dose of radiation was Gy in 25 fractions (16). Adjuvant systemic therapy. Before 2000, node-positive patients were treated with one of four chemotherapy regimens AC (doxoru-

3 Prediction of LRR after mastectomy S. H. CHENG et al bicin, cyclophosphamide), CMF (cyclophosphamide, methotrexate, 5-fluorouracil), CAF (cyclophosphamide, doxorubicin, 5-fluorouracil), or CEF (cyclophosphamide, epirubicin, 5-fluorouracil) as were selected node-negative patients (those who were ER/progesterone receptor negative or premenopausal patients with a 2-cm primary lesion). Patients with 4 or more positive lymph nodes during this period received A followed by CMF (17). After 2000, nodepositive patients with ER negative were treated with AC followed by paclitaxel; node-positive patients with ER positive were treated with A followed by CMF. Patients with 10 or more positive nodes received dose-dense ACT. Prognostic factors Fifteen tumor-, patient-, and treatment-related factors were examined for their impact on LRR (Table 1). We included all characteristics that have been previously reported in the literature to influence LRR, regardless of the strength of the evidence. These factors might be characterized as patient related (e.g., age), tumor related (e.g., ER status), or treatment related (e.g., adjuvant chemotherapy). Of the 1,143 patients who underwent modified radical mastectomy during the study, 1,010 patients who have data on all six clinical risk factors (age at diagnosis, primary tumor size, axillary lymph node status, nuclear grade, LVI, hormonal receptor status) constitute the study group. Statistical analysis Two statistical approaches were used in this study; these approaches complement each other by using different types of mathematical models (linear and nonlinear) to characterize the relationship between LR control and the prognostic factors under study. In the first approach, Cox proportional hazards regression models were used to assess the prognostic significance of risk factors (14). All patients were stratified by lymph node involvement ( 4 vs. 4) and LR status (recurrence vs. control), and then, within strata, patients were randomly assigned to one of two subsets of equal size for model training and validation. Patient characteristics in the training and test data sets were well balanced (Table 1). The step-down regression analysis of LR control in the training samples evaluated all 15 prognostic variables mentioned above. Duration of LR control was calculated from the first day of treatment until the day when chest wall or regional lymph node recurrence was observed. Local regional control estimates were calculated according to the methods of Kaplan and Meier (18). Log rank tests were used to assess the statistical significance of the difference in the probabilities of LR control between specific patient subsets (19). The model results were used to define levels of risk of LR recurrence (see below); the predictive ability of these levels was assessed by application to the validation set. The Kaplan-Meier estimates of LR control in specific patient subsets were compared with results from the second statistical approach, based on prediction tree models. These tree models classify patients into subgroups with higher or lower probability of LR control, according to selected risk factors (e.g., lymph node positive 3 vs. 3, age 40 vs. 40 years). Estimates of LR control were calculated for each subgroup. The details of tree generation and Bayesian analysis have been published previously (13, 20, 21). In brief, Bayesian methods of analysis are used to fit candidate tree models to the training data; each model uses different risk factors to assign patients to subgroups, which vary in their probability of LR control. The bestfitting trees are selected, and from each of these trees an estimate of LR control for each patient is calculated. This provides several estimates of LR control for each patient; the final estimate of LR control for a patient is a weighted average of these estimates, whereby the estimates from better-fitting trees are given larger weights. The risk factors used to subgroup patients in the bestfitting models provide insights into which clinical predictor variables and interactions of clinical predictor variables impact the probability of LR control. We selected 255 of the 1,010 patients for training and construction of the prediction tree models. The selection criteria were patients who had any evidence of LR recurrence within 5 years (n 53) and no evidence of LR recurrence without PMRT for a minimum follow-up of 5 years (n 202). The predictive ability of the model to assess probability of LR control is assessed by cross-validation on the training set. RESULTS Table 1 shows the clinical characteristics and treatment parameters for the 1,010 patients in the study group, divided equally into training and test subsets. These subsets were well balanced with regard to these risk factors. In general, the patients were young, with 23% aged 40 years at the time of diagnosis. Despite this, 71% were ER positive. Forty-nine percent were node negative, and 45% had primary tumors 2 cm in size or less. Regarding treatment characteristics, 72% of patients received adjuvant chemotherapy, 72% adjuvant hormonal therapy, and 26% adjuvant radiotherapy. Chemotherapy regimens consisted of CMF in 19.7%, CAF in 38.3%, CEF in 1.6%, AC in 12%, A-CMF in 14.5%, ACT in 2.6%, dose-dense ACT in 4.7%, and miscellaneous in 6.6%. The radiation dose ranged from 45 to 66 Gy (median, 50 Gy). With a median follow-up of 48 months, 869 patients (86%) were without evidence of disease, whereas 141 (14%) had relapsed. The initial sites of relapse were local regional in 1.6% of patients, distant in 8%, and both combined in 4%. The 5-year probability of any LRR, either isolated or in combination with distant metastases, was 7.2%. Multivariate analysis of both the training data set and the test set revealed four clinical characteristics significantly Table 2. Cox proportional hazards analysis of risk of 5-year local regional recurrence in the training data set Risk factor Hazard ratio 95% CI Age (y) 40 vs Estrogen receptor status Negative vs. positive Lymph node positive 1 3/4 9/ 9 vs /16.1/ /3.4 76/ Lymphovascular invasion Prominent vs. absent/focal Adjuvant radiotherapy Yes vs. no Abbreviation: CI confidence interval. n 506.

4 1404 I. J. Radiation Oncology Biology Physics Volume 64, Number 5, 2006 associated with LRR. These were axillary nodal status, ER status, LVI, and age at diagnosis. In addition, the use of adjuvant radiotherapy was significantly associated with LR control. The best grouping of these clinical risk factors was age ( 40 vs. 40 years), ER (negative or positive), axillary nodal involvement (0 vs. 1 3 vs. 4 9 vs. 9), and LVI (absent/focal vs. prominent). Hazard ratios for these risk factors are shown in Table 2. The four pretreatment characteristics were used to design a model to predict an individual patient s risk of LRR. The relative risks associated with each of the risk factors were comparable (hazard ratios of ) with the exception of patients with 4 9 positive nodes or 9 positive nodes, with hazard ratios of 16.1 and 45.8, respectively. A prognostic index score was then developed for the risk of LRR. The index score was defined as the sum of the number of risk factors present with each risk factor receiving a value of 1, except for lymph node status, which was scored as 1 for 1 3 nodes positive, 2 for 4 9 nodes positive, and 3 for 9 nodes positive. Thus, patients could receive a score from 0 to 6. Risk groups were defined by comparing the relative risk of LRR in patients with different index scores and combining categories with similar relative risks. Thus, patients could be grouped as low risk (score 0 or 1), intermediate risk (score 2 or 3), or high risk (score 4 6). These three groups showed distinctive differences, not only in the risk of LRR but also in terms of metastasis-free survival and overall survival (Figs. 1a c, Table 3). The low-risk group (score 0 or 1) comprised 547 patients, 522 of whom did not receive PMRT and 25 of whom did. There were no significant differences in outcome whether PMRT was given or not. In the intermediate-risk group (score 2 or 3) there were 328 patients, 126 of whom received PMRT, 202 of whom did not. Although LRR was significantly reduced in patients receiving PMRT (p ), there were no significant differences in metastasis-free survival or overall survival. In the high-risk group (score 4 6) there were 135 patients, 109 of whom received PMRT, 26 of whom did not. Postmastectomy radiotherapy significantly improved Fig. 1. Probability curves for (a) local regional (LR) control, (b) metastasis-free survival, and (c) overall survival, based on the prognostic index score, which divided patients into low-, intermediate-, and high-risk groups. CI confidence interval.

5 Prediction of LRR after mastectomy S. H. CHENG et al Table 3. Five-year local regional control and survival Group No. of patients 5-y LR control probability (%) 5-y metastasis-free probability (%) 5-y overall survival (%) Cox model Low-risk patients (score 0 1) No PMRT PMRT Total Intermediate-risk patients (score 2 3) No PMRT PMRT Total High-risk patients (score 4 6) No PMRT PMRT Total Tree model Low-risk patients (predictive index 0.94) No PMRT PMRT Total Intermediate-risk patients (predictive index ) No PMRT PMRT Total High-risk patients (predictive index 0.71) No PMRT PMRT Total Abbreviations: LR local regional; PMRT postmastectomy radiotherapy. LR control, metastasis-free survival, and overall survival (p , 0.001, and , respectively.) Bayesian tree analysis To provide an initial indication of robustness and accuracy, we evaluated the predictive probability of LR control for individual patients of the 255-patient training data set, using leave-one-out cross-validation; the tree model process was recomputed repeatedly, each time leaving out one sample and then predicting it on the basis of the rest. The prediction tree methodology generated six best-fitting tree models (Fig. 2). These models identified 5 of the 15 candidate risk factors as important (lymph node status, age at diagnosis, ER status, LVI, and primary tumor size). Three major tree models (Trees 2, 43, and 13, weighting 42%, 33%, and 16%, respectively) are presented in Figs. 3A C. The estimates of 5-year LR control from these trees were averaged to produce overall estimates of 5-year LR control for each patient (the tree numbers are automatically generated by MATLAB software; MathWorks, Natick, MA). As shown in Fig. 3, axillary lymph node involvement is the first risk factor selected to split patients into high- and low-risk subgroups, followed by age at diagnosis and ER status. Patients with a prior condition of lymph node negative and age 35 years in Tree 13 are estimated to have a 5-year LR control probability of 79%. In contrast, patients with 1 3 positive lymph nodes, age 38 years, and ER Fig. 2. Summaries of predictor variables and the clinical tree model: five clinical variables (age at diagnosis, primary tumor size, lymph node status [lnpos], estrogen status [erlevel], and lymphovascular invasion [lvi]) (columns) that appear in the selected top trees (rows), and the levels (boxed numbers) of the trees in which they define node splits. The probability of each tree and the overall probability of occurrence of each of the clinical factors across the set of trees are also given. Weighted average of estimates from individual trees are shown in parentheses in the column headings; utility estimates of clinical risk factors in tree models are shown in parentheses in the row headings.

6 1406 I. J. Radiation Oncology Biology Physics Volume 64, Number 5, 2006 Fig. 3. Three most common trees: the boxes at nodes of a tree indicate the number of patients, whereas the number next to the box represents the model-based estimate of 5-year LR control survival probability. (a) Tree 2 is the most important tree of our tree models, weighted 42% (see Fig. 2); the first node is number of positive axillary lymph nodes, with a cutoff value of 3; the second nodes are age 38 years and estrogen receptor moderately positive; the third level of split is axillary lymph node positive, followed by estrogen receptor moderately positive, and so on. (b) Tree 43 is the second most important tree, weighted 33%; similar to Tree 2, patients initially are split by number of positive axillary lymph nodes positive, with a cutoff value of 3; age 38 years and estrogen receptor strongly positive are the second nodes; lymph node negative and estrogen receptor moderately positive are the third nodes, and so on. (c) Tree 13 is the third most important decision tree, consisting of 16% of our prediction tree models, which means of lower weight in the prediction of recurrence. Patients with a prior condition of lymph node negative and age 35 years would have a 5-year probability of LR control of 79%. Abbreviations: ER status: negative ( ); node weakly positive ( ); node densely positive ( ); node strongly positive ( ).

7 Prediction of LRR after mastectomy S. H. CHENG et al Fig. 3. (Continued) moderately positive in Tree 43 have excellent 5-year estimates of LR control, at 95%. Patients with good LR control generally had a high predictive index, and patients with LR recurrence had a low predictive index. Both prediction patterns had small prediction uncertainty. As compared with Cox proportional hazards models, the tree models focuses on predicting individual risk of LR recurrence. The computer program to estimate individual risk is available at cpo/cpo.php. We applied the prediction tree models to all 1,010 patients to obtain a predictive index for each individual patient. Comparing Bayesian classification tree models with Cox proportional hazards model A grouping of the 1,010 patients based on their individual predictive index is shown in Table 3. A predictive index of 0.71 (high-risk group) is probably the best cutoff value for patients to undergo PMRT; PMRT in this group improved not only LR control but also metastasis-free and overall survival. In contrast, patients did not obtain benefit from PMRT if their predictive indices were For patients with predictive indices between 0.71 and 0.94, PMRT improved LR control (from 89.4% to 100%, p 0.005) but made no difference in metastasis-free or overall survival. Multivariate analysis of treatment effects in the high-risk groups defined by either prognostic index score or prediction tree models revealed that PMRT, adjuvant chemotherapy, and hormonal therapy each significantly improved LR control. Postmastectomy radiotherapy also significantly improved metastasis-free survival and overall survival, whereas the improvement in metastasis-free survival and overall survival due to either adjuvant chemotherapy or hormonal therapy is evident but not consistently significant (Table 4). DISCUSSION The role of PMRT in the management of breast carcinoma has been and is controversial since its introduction more than 5 decades ago. It was soon established that PMRT significantly decreased LRR after surgery. Innumerable phase II studies and more than 30 randomized trials have shown a consistent relative risk reduction in LRR (of two thirds to three quarters), as summarized in three recent meta-analyses (22 24). The absolute benefit depends, of course, on the relative risk of an LRR, being greatest in those patients at highest risk for LRR. Survival benefits obtained from PMRT are the center of the controversy. Many early trials failed to show a survival benefit. The complexities of these trials and the pitfalls therein have been extensively reviewed by Recht and Edge (25). Three major recent trials, however, have convincingly shown survival benefits, as has a recent meta-analysis (6 8, 22). Patients have generally been selected for PMRT (and inclusion in these trials) on the basis of nodal involvement, traditionally grouped as 1 3 nodes involved or 4 nodes positive. Increasing number of involved lymph nodes generally correlates with both the risk of systemic

8 1408 I. J. Radiation Oncology Biology Physics Volume 64, Number 5, 2006 Table 4. Multivariate analysis for treatment effects on local regional recurrence/metastasis-free/ overall survival in high-risk patients Hazard ratio (95% CI) Treatment factors LR recurrence Metastasis-free survival Overall survival By prognostic index model (n 135) Adjuvant radiotherapy 0.16 ( ) 0.4 ( ) 0.36 ( ) Adjuvant chemotherapy 0.11 ( ) 0.43 ( ) 0.57 ( ) Adjuvant hormonal therapy 0.34 ( ) 0.85 ( ) 0.59 ( ) By prediction tree models (n 206) Adjuvant radiotherapy 0.10 ( ) 0.47 ( ) 0.58 ( ) Adjuvant chemotherapy 0.07 ( ) 0.25 ( ) 0.37 ( ) Adjuvant hormonal therapy 0.24 ( ) 0.68 ( ) 0.49 ( ) Abbreviation as in Table 2. spread as well as that of LRR. The three trials mentioned above, though comprising mostly patients with 1 3 positive nodes, either showed a greater benefit for the 4 node-positive group or were insufficiently powered for subset analysis (26). An expert American Society of Clinical Oncology (ASCO) panel developing guidelines for PMRT concluded that although there was strong evidence for increased local control in both nodal groupings, the evidence for improvement in disease-free survival and overall survival was more consistent for those with 4 nodes involved, compared with the 1 3 nodepositive group. The panel concluded, the evidence was insufficient to make recommendations for the routine use of PMRT in patients with 1 3 positive nodes (5). The panel also addressed the influence of other tumorrelated characteristics (e.g., grade, ER status, LVI, primary size) as well as patient-related factors (e.g., age, menopausal status) and concluded that evidence was insufficient to know how these variables should factor into the decision of whether to use PMRT. Further investigations in these areas were strongly recommended. Unfortunately, evidence from phase III trials on some of these points will not be soon forthcoming. A North American Breast Intergroup study looking specifically at patients with 1 3 positive lymph nodes and randomizing them to PMRT or not has closed owing to poor accrual. A European study is planned but is years from completion (2). Given this background, we set out to develop a prognostic index to guide decisions regarding the use of PMRT that might be more informative than a simple division into 1 3 or 4 lymph nodes involved. Using Cox proportional hazards analysis, we identified patient age, ER status, and LVI, in addition to nodal status, as important prognostic variables and were able to group patients into three distinct groups with low, intermediate, and high risks of LRR. Although PMRT increased local regional control in all groups, statistically significant survival benefits were only seen in the high-risk group (Table 4). The great majority of patients with 4 nodes involved fall into this high-risk group, but so do significant numbers of patients with 1 3 positive nodes (e.g., a young woman with ER-negative disease with LVI and 1 3 nodes involved would fall into the high-risk group). The model does suggest that PMRT is not necessary for patients with 1 3 lymph nodes involved in the absence of other risk factors. The predictive tree model gave similar results. Only the highest-risk group, with a predictive index of 0.71, would experience survival benefits from PMRT, although all groups would have decreased LRR. These results are consistent with recent calculations performed by Olivotto et al. (27), suggesting a ratio between the number of LRRs prevented and survival of approximately 4 to 5 to 1. Thus, an absolute reduction in LRR from 15%, for example, to 3% would be expected to lead to an absolute survival benefit of 2% to 3%. A demonstration of statistically significant survival improvement would require thousands of patients in trials, similar to what has been entered in trials of adjuvant systemic therapy. In patients with a much higher risk for LRR, the absolute benefit is correspondingly greater, and larger absolute survival benefits are expected as well. The ASCO expert panel concluded that, similar to systemic therapy, it was highly likely that PMRT achieved proportional reductions in the risk of LRR that were the same for all treated groups, the absolute benefit depending on the relative risk of the event, as previously stated. In summary, we have developed a prognostic score and predictive index for LRR based on easily obtainable clinical information that seems to provide more information regarding the risk of LRR than a simple division of patients by primary size and the number of lymph nodes involved. This prognostic score and predictive index might lead to improved guidelines for selection of patients for PMRT (Table 5). Confirmation of these results by other groups is highly desirable. As with adjuvant systemic therapy, the ultimate decision as to whether to undergo PMRT rests with the patient and relates to what level of risk she finds acceptable and how much treatment she is willing to undertake for a small survival benefit.

9 Prediction of LRR after mastectomy S. H. CHENG et al Table 5. Impact of postmastectomy radiotherapy on patient outcomes, according to prognostic score and predictive index Prognostic score or predictive index Improvement of LR control Improvement of metastasis-free survival Improvement of overall survival 2/ 0.94 No No No 2 3/ Yes No No 3/ 0.71 Yes Yes Yes Abbreviation as in Table 3. REFERENCES 1. Griem KL, Henderson IC, Gelman R, et al. The 5-year results of a randomized trial of adjuvant radiation therapy after chemotherapy in breast cancer patients treated with mastectomy. J Clin Oncol 1987;5: Whelan T, Levine M. More evidence that locoregional radiation therapy improves survival: What should we do? J Natl Cancer Inst 2005;97: Overgaard M, Christensen JJ, Johansen H, et al. Evaluation of radiotherapy in high-risk breast cancer patients: Report from the Danish Breast Cancer Cooperative Group (DBCG 82) Trial. Int J Radiat Oncol Biol Phys 1990;19: Fowble B, Gray R, Gilchrist K, et al. Identification of a subgroup of patients with breast cancer and histologically positive axillary nodes receiving adjuvant chemotherapy who may benefit from postoperative radiotherapy. J Clin Oncol 1988;6: Recht A, Edge SB, Solin LJ, et al. Postmastectomy radiotherapy: Clinical practice guidelines of the American Society of Clinical Oncology. J Clin Oncol 2001;19: Overgaard M, Hansen PS, Overgaard J, et al. Postoperative radiotherapy in high-risk premenopausal women with breast cancer who receive adjuvant chemotherapy. Danish Breast Cancer Cooperative Group 82b Trial. N Engl J Med 1997;337: Overgaard M, Jensen MB, Overgaard J, et al. Postoperative radiotherapy in high-risk postmenopausal breast-cancer patients given adjuvant tamoxifen: Danish Breast Cancer Cooperative Group DBCG 82c randomised trial. Lancet 1999;353: Ragaz J, Jackson SM, Le N, et al. Adjuvant radiotherapy and chemotherapy in node-positive premenopausal women with breast cancer. N Engl J Med 1997;337: Dunst J, Steil B, Furch S, et al. Prognostic significance of local recurrence in breast cancer after postmastectomy radiotherapy. Strahlenther Onkol 2001;177: Katz A, Buchholz TA, Thames H, et al. Recursive partitioning analysis of locoregional recurrence patterns following mastectomy: Implications for adjuvant irradiation. Int J Radiat Oncol Biol Phys 2001;50: Sundquist M, Thorstenson S, Klintenberg C, et al. Indicators of loco-regional recurrence in breast cancer. The South East Swedish Breast Cancer Group. Eur J Surg Oncol 2000;26: A predictive model for aggressive non-hodgkin s lymphoma. The International Non-Hodgkin s Lymphoma Prognostic Factors Project. N Engl J Med 1993;329: Pittman J, Huang E, Nevins J, et al. Bayesian analysis of binary prediction tree models for retrospectively sampled outcomes. Biostatistics 2004;5: Cox D. Regression models and life-tables. J R Stat Soc 1972; 34: Huang E, Cheng SH, Dressman H, et al. Gene expression predictors of breast cancer outcomes. Lancet 2003;361: Cheng SH, Jian JJ, Chan KY, et al. The benefit and risk of postmastectomy radiation therapy in patients with high-risk breast cancer. Am J Clin Oncol 1998;21: Cheng SH, Tsou MH, Liu MC, et al. Unique features of breast cancer in Taiwan. Breast Cancer Res Treat 2000;63: Kaplan EL, Meier P. Nonparametric estimation from incomplete observation. J Am Stat Assoc 1958;53: Mantel N, Haenszel W. Chi-square test with one degree of freedom: Extensions of the Mantel-Haenszel procedure. JAm Stat Soc 1963;58: Pittman J, Huang E, Dressman H, et al. Integrated modeling of clinical and gene expression information for personalized prediction of disease outcomes. Proc Natl Acad Sci U S A 2004; 101: Kass R, Raftery A. Bayes factors. J Am Stat Assoc 1995;90: Whelan TJ, Julian J, Wright J, et al. Does locoregional radiation therapy improve survival in breast cancer? A metaanalysis. J Clin Oncol 2000;18: Favourable and unfavourable effects on long-term survival of radiotherapy for early breast cancer: An overview of the randomised trials. Early Breast Cancer Trialists Collaborative Group. Lancet 2000;355: Cuzick J, Stewart H, Rutqvist L, et al. Cause-specific mortality in long-term survivors of breast cancer who participated in trials of radiotherapy. J Clin Oncol 1994;12: Recht A, Edge SB. Evidence-based indications for postmastectomy irradiation. Surg Clin North Am 2003;83: Ragaz J, Olivotto IA, Spinelli JJ, et al. Locoregional radiation therapy in patients with high-risk breast cancer receiving adjuvant chemotherapy: 20-year results of the British Columbia randomized trial. J Natl Cancer Inst 2005;97: Olivotto IA, Truong PT, Chua B. Postmastectomy radiation therapy: Who needs it? J Clin Oncol 2004;22: